On the Prevalence and Appropriate Use of Data-Model-Based Informative Priors for Estimating Multilevel Linear Models

• In Event: Statistical Frontiers: Bayesian Inference, Machine Learning, and Longitudinal Modeling

Abstract

Bayesian analysis has been on the rise, yet the issue of prior specification continues to pose challenges. In this paper we investigate the prevalence of Bayesian analyses and prior specification choices as well as the consequences of using data-based priors on fixed effect parameter recovery for linear random intercept models. Our systemic review results showed that researchers have largely relied on software default priors – and on rare occasions they specify informative priors – one-third use sample-related values. Our simulation results showed that posterior standard errors are progressively underestimated as data-based prior informativeness increases. When using data-based priors, we found that the best approach for obtaining unbiased coefficient credible intervals is to use weakly informative priors or a random split-half method.

Authors

• Zhigang Zhang, University of Washington

• Elizabeth A. Sanders, University of Washington

• Timothy R. Konold, University of Virginia